Nathan Lambert's Harvard lecture video released: Chinese open-source models lead, US compute advantage about 16x

Nathan Lambert's Harvard lecture: a panoramic view of the open-source model landscape—China leads across the board in open source, but the US has a compute advantage of more than 10x; distillation is not the key issue In his guest lecture at Harvard University's CS 2881R "AI Security," @natolambert argues: open-source models lag the closed-source frontier by roughly 3 to 5 months, and the gap is still narrowing. Across performance, downloads, and paper citations, Chinese open-source models lead comprehensively, while US open-source models lag China by another 3 to 4 months. There are three core judgments: treating "distillation" as the core issue overestimates it; the US's truly durable advantage is compute; and banning open source would only harm economic activity without stopping bad actors. https://t.co/GpErvGTR0w What counts as an open-source model—the license is key · Open source is a spectrum: from weight-only releases (Llama, Qwen, DeepSeek, Kimi, GLM) to fully open source with data and code (OLMo, Pythia, SmolLM, Marin), with Nemotron in between. · Permissive licenses: MIT (DeepSeek, GLM-5.2, MiMo), Apache 2.0 (Qwen, gpt-oss, Gemma 4). · Custom licenses with additional conditions are not really open source: revenue thresholds (Kimi K3, Muse Spark Open, and MiniMax M3 have an annual revenue threshold of $20 million; GLM-5.3 requires a security review above $10 billion), branding requirements, usage restrictions. Most of these clauses have not yet been tested in court. Why open source Support for open source comes from several directions: capital (turning complements into cheap commodities), research and education, preventing concentration of power, and freedom. In Bill Gurley's phrase, open source is also "the natural choice for the second place," used to capture market and distribution channels and gain soft power. Chinese labs collectively moving toward open source is about gaining visibility in Western AI circles; after DeepSeek R1, Xiaomi, Meituan, Ant, and others also entered the fray. Enterprises use open-source models to "own their own models"—for example, Harvey did legal-domain post-training on Kimi K3, nearly doubling its overall pass rate at essentially unchanged cost. Competitiveness · In 2026, anything under 1T parameters counts as a "small model." Recent strong models: Xiaomi MiMo-V2.6-Pro (1T total parameters, 42B activated, Artificial Analysis index 46), DeepSeek V4.1 Flash (763B, 45), MiniMax M3.1 Flash (42, with third-party evaluations below official claims). · On "benchmark gaming": Chinese models do place more emphasis on public leaderboards, and their long-tail capabilities are less comprehensive than US models. But he spent half a year and could not find any analysis demonstrating clear flaws in the leading open-source models (Kimi, GLM, DeepSeek). · Open-source models still anchor the optimal price-performance range, though closed-source vendors are now competing in that range too. The distillation debate (the most opinionated part of the whole piece) · Training has three stages: pretraining, mid-training and SFT, and reinforcement learning. Distillation mainly happens in the SFT stage, and reinforcement learning now consumes more compute than pretraining. · What is called "distillation" today is actually adversarial distillation: finding ways to extract hidden reasoning traces from APIs. Many Chinese labs have dedicated teams doing this. · He argues it is not worth making this the focus: public research has not shown that SFT on reasoning traces yields huge gains; there is no good policy tool that targets China without harming the global open-source ecosystem; and Chinese labs are strong in their own right—"fast-following" is inherently more compute-efficient than pioneering the frontier. Xiaomi publicly livestreamed the reinforcement learning training of MiMo-V2.6 (about 1.5 million rollouts, costing $3.47 million), which is evidence of not relying solely on distillation. · The truly durable advantage is compute: the total compute of the top five US labs averages 16 times that of the top five Chinese labs; by the end of 2026, Anthropic and OpenAI will each have more than 5 GW of compute, while the top Chinese labs will have at most 100 to 200 MW. Usage Inference volume for open-source models is growing rapidly, and local model downloads are accelerating; Hugging Face has about 600,000 daily unique users. When performance is similar, adoption of US models is 2 to 5 times that of Chinese models, indicating market demand for US open-source models. He scanned 370,000 arXiv papers, and Qwen is now mentioned nearly as often as OpenAI's GPT. The summary: China leads in performance, downloads, and research, and the inference and post-training markets are also dominated by Chinese models. Risk and governance · AI-driven cyberattacks are already happening in reality, and the publicly known cases mainly involve abuse of closed-source models (Claude Code, Codex). · Banning open-source models is unfeasible in practice (the US cannot build an effective "firewall"), and more than a dozen labs can train models at roughly the frontier level of about a year ago; the result would only be to suppress economic activity. Many claims about open-source risks are exaggerated (such as disinformation), but risks like deepfakes are indeed being accelerated. · Recommendation: releasers should assess the "additional risk" of each model individually, with national unified risk standards, rather than leaving it to each company to judge for itself. Open-source models are critical for cyber defense, because companies' trusted access programs cannot cover the entire economy.

View original

Nathan Lambert 哈佛讲座：开源模型现状全景，中国开源全面领先，但美国有 10 倍以上的算力优势；蒸馏不是关键问题 @natolambert 在哈佛大学 CS 2881R「AI 安全」的客座讲座，他认为：开源模型落后闭源前沿大约 3 到 5 个月，差距还在缩小。性能、下载量、论文引用这三方面，中国开源模型全面领先，美国开源模型又比中国落后 3 到 4 个月。核心判断有三点：把"蒸馏"当成核心问题是高估了；美国真正持久的优势是算力；禁止开源只会损害经济活动，挡不住坏人。 https://t.co/GpErvGTR0w 什么是开源模型，许可证是关键 · 开源是一个光谱：从只开放权重（Llama、Qwen、DeepSeek、Kimi、GLM），到连数据和代码都开放的完全开源（OLMo、Pythia、SmolLM、Marin），Nemotron 处在中间。 · 宽松许可证：MIT（DeepSeek、GLM-5.2、MiMo），Apache 2.0（Qwen、gpt-oss、Gemma 4）。 · 自定义许可证带附加条件，并不算开源：营收门槛（Kimi K3、Muse Spark Open、MiniMax M3 的年营收门槛是 2000 万美元，GLM-5.3 超过 100 亿美元要做安全审查）、品牌露出要求、使用限制。这些条款大多还没有在法庭上检验过。 为什么要开源 支持开源的力量来自几个方向：资本（把互补品变成廉价商品）、科研和教育、防止权力集中、自由。按 Bill Gurley 的说法，开源也是"第二名的自然选择"，用来抢市场和分发渠道、获得软实力。中国实验室集体走向开源，是为了在西方 AI 圈获得存在感；DeepSeek R1 之后，小米、美团、蚂蚁等也纷纷入场。企业用开源模型来"拥有自己的模型"，例如 Harvey 在 Kimi K3 上做法律方向的后训练，全部通过率接近翻倍，成本基本不变。 竞争力 · 2026 年，1T 参数以下就算"小模型"了。新近的强模型：小米 MiMo-V2.6-Pro（总参数 1T、激活 42B，Artificial Analysis 指数 46）、DeepSeek V4.1 Flash（763B，45）、MiniMax M3.1 Flash（42，第三方测评低于官方宣传）。 · 关于"刷榜"：中国模型确实更看重公开榜单，长尾能力也不如美国模型全面。但他花了半年时间，也没找到能证明头部开源模型（Kimi、GLM、DeepSeek）有明显缺陷的分析。 · 开源模型依然锚定着性价比的最优区间，不过闭源厂商现在也开始在这个区间竞争。 蒸馏之争（全篇最有观点的部分） · 训练分三个阶段：预训练、中期训练和 SFT、强化学习。蒸馏主要发生在 SFT 阶段，而现在强化学习消耗的算力已经超过预训练。 · 现在说的"蒸馏"其实是对抗性蒸馏：想办法从 API 里套取被隐藏的推理过程。中国不少实验室有专门的团队做这件事。 · 他认为不值得把它当成焦点：公开研究并没有显示用推理过程做 SFT 能带来巨大提升；没有好的政策手段可以只针对中国而不伤害全球开源生态；中国实验室本身就很强，"快速跟进"本来就比开拓前沿省算力。小米公开直播了 MiMo-V2.6 的强化学习训练（约 150 万次 rollout，花费 347 万美元），就是不只依赖蒸馏的证据。 · 真正持久的优势是算力：美国前五家实验室的总算力平均是中国前五家的 16 倍；到 2026 年底，Anthropic 和 OpenAI 各有 5 GW 以上的算力，而中国头部实验室最多只有 100 到 200 MW。 使用情况 开源模型的推理量快速增长，本地模型下载在加速；Hugging Face 日均独立用户约 60 万。性能相近时，美国模型的采用量是中国模型的 2 到 5 倍，说明市场对美国开源模型有需求。他扫描了 37 万篇 arXiv 论文，Qwen 被提到的次数已经接近 OpenAI 的 GPT。总结是：性能、下载、科研三方面中国都领先，推理和后训练市场也主要是中国模型。 风险与治理 · AI 驱动的网络攻击已经真实发生，而目前公开的案例主要是闭源模型（Claude Code、Codex）被滥用。 · 想禁止开源模型，在现实中做不到（美国不可能建起有效的"防火墙"），而且十多家实验室都能在前沿一年左右的水平上训练模型，结果只会抑制经济活动。不少关于开源风险的说法被夸大了（比如虚假信息），但深度伪造这类风险确实被加速了。 · 建议：发布方按每个模型评估它带来的"额外风险"，并由国家统一风险标准，而不是让各家公司自己判断。开源模型对网络防御非常关键，因为公司的受信任访问项目没法覆盖整个经济。